Age-related determinants of geriatric depression: an analysis of theory-driven and data-driven approaches

BACKGROUND: Geriatric depression is a prevalent mental health condition whose risk profile may vary across age strata. This study examined whether depression risk and protective factors differ across age groups (65-74, 75-84, 85+) among community-dwelling older adults in Turkey. METHODS: Data were derived from the Turkey Elderly Profile Survey 2023 (TEPA-2023), a nationally representative multi-stage cluster sample. Among 29,785 individuals surveyed, the final analytical sample comprised 8,370 adults aged 65 years and older. Depression was defined using the Geriatric Depression Scale-30 (GDS-30) with a cut-off score of ≥ 11. In the theory-driven phase, binary logistic regression models including 24 independent variables were estimated for the overall sample and for separate age groups, with model validation based on stratified 10-fold cross-validation. In the data-driven phase, multiple machine learning algorithms were compared using accuracy, recall, F1-score, and area under the ROC curve (AUC). RESULTS: The prevalence of depression increased significantly with age, from 39.9% in the 65-74 group to 53.4% in the 75-84 group and 67.7% in the 85 + group (χ² = 242.40, p < 0.001). In the overall logistic regression model, the strongest risk factor was unhappiness (OR = 2.516), whereas the strongest protective factor was functional independence as measured by Lawton-Brody IADL (OR = 0.386). The overall model showed good discrimination (AUC = 0.812; McFadden R² = 0.245). Age-stratified analyses showed that the number of significant predictors decreased with advancing age (13→9→5), and obesity showed a protective effect only in the 85 + group (OR = 0.506, p = 0.016). In the data-driven phase, ten machine learning algorithms were compared using stratified 10-fold cross-validation. Gradient Boosting achieved the highest overall AUC (0.813), while SVM-RBF yielded the best discrimination in the 75-84 (0.806) and 85+ (0.828) groups. Permutation importance analysis converged with logistic regression findings, consistently ranking unhappiness, self-rated health, Lawton-Brody IADL, and physical activity frequency as top predictors. Across age strata, predictor prominence shifted from education and income in the Under-75 group toward BMI and disability indicators in the 85 + group. CONCLUSIONS: The risk profile of geriatric depression changes substantially with advancing age. While socioeconomic factors appear more prominent in younger-old adults, health status and functional independence become more dominant in later old age. The combined use of theory-driven regression and data-driven machine learning provides a more comprehensive framework for identifying age-related depression patterns in older adults.